Research ·

Knowledge search success and article reuse in support queues

A knowledge article only helps when an agent can find it during a live ticket, so search success should be judged by the answer, not by the click.

Knowledge search success and article reuse in support queues research hero

Key stats

3

failed-search patterns reviewed each week

2

levels of measurement: article and queue

1

content owner for the search vocabulary

Methodology and findings

Question and scope

This review asks how an outsourced help desk can measure whether agents find and reuse the right knowledge article without turning search data into surveillance. The population is searches performed inside the support tool and the articles opened from the results. The unit is one search session tied to a ticket. The question matters because an article that exists but cannot be found costs the same as an article that was never written. It treats search as part of the answer path, not as a separate content system.

The scope covers the desk's own knowledge base and the search interface agents use during a ticket. It excludes public help-centre search, which serves a different audience with different language. The review does not claim that a click proves understanding or that a zero-result search proves a missing article. It describes measures that can guide content work and a search review. It also excludes searches by customers on the public site, where the language and the intent differ.

What search success means

Search success means the agent found a page that answered the question and used it in the reply. A click is only a step toward that outcome. A session can end with a click and still fail because the article was wrong, outdated, or about a different plan. Measure the outcome, not only the interaction. Record the article used in the ticket so the link between the search and the answer can be checked later.

Define success per request type, because some questions need a procedure and others need a policy statement. A search that finds a policy page for a procedural question is a near miss. Record the request type beside the search so the analysis can separate those cases. A single success rate across all types hides the pattern. Use the same request types as the routing taxonomy so the two reports can be read together.

The difference between a click and an answer

Pair search data with ticket outcomes: was the ticket resolved without a second contact, did the customer reopen it, did the agent edit the saved reply. These signals are imperfect, but together they say more than a click count. A page that is opened often and followed by a reopen is a candidate for review. Review the sample with the content owner, because the pattern usually needs a content change rather than coaching.

Read a small sample of sessions and compare the article with the customer's wording. The reviewer should ask whether the article answered the question the customer actually asked. This is a judgment, not a metric, and it should be recorded with an example. The sample keeps the dashboard from becoming the only version of the truth. Keep the examples anonymous and focus the note on the article and the question, not on the agent.

Instrumenting failed searches

Capture zero-result searches, searches that lead to no click, and searches followed by a second search with different words. Each points to a different problem: a missing article, a title that does not match, or a result list that buries the right page. Count them by request type rather than in total. Store the exact query text, because a normalized version can hide the words that failed.

Review the top failed terms weekly and group them by intent. A term that appears across many tickets may deserve an article, while a one-off misspelling may need a synonym. Do not create an article for every failed search, because a bad question is not always a content gap. The grouping is what separates a pattern from noise. Assign each group to the desk or the client, depending on who owns the page that should answer it.

Titles and customer vocabulary

Write titles with the words customers and agents use, not the internal name of the feature. A title that begins with the product name may sort well and still fail in search. Put the task or symptom first, then the product context. Test the title against the last twenty tickets that asked the question. Check the title in the result list, where it competes with other pages rather than standing alone.

Add synonyms and common misspellings where the platform supports them, and record why each was added. A synonym list without an owner grows into a pile of guesses. Review it when the product vocabulary changes. The goal is to match the words in the queue, not to build a dictionary. Remove a synonym when the product term changes, so the list does not preserve an old name.

Measuring reuse without surveillance

Measure at the article and queue level, not at the individual agent level. Report which articles are opened, which are followed by edits, and which are never opened. Aggregate data can guide content work without ranking people. If individual data is needed for coaching, keep it in the quality process with a stated purpose. Publish the purpose of the measurement to the team before the first report.

Tell agents what is measured and why, and give them a way to flag a bad article from the search result. A flag is more useful than a click because it carries a reason. Review flags with the content owner and close them with a change or a decision. Trust in the measurement depends on how it is used. Close the loop on flags with a visible change or a written reason, or agents will stop raising them.

When an article should not be reused

Some situations require a fresh answer even when an article exists: a safety report, a legal request, or a case with facts the article does not cover. Teach agents to recognize the boundary and to escalate instead of forcing a template. A reused article in the wrong case can be worse than no article, because it sounds authoritative. Put the exceptions in the article itself so an agent sees the boundary at the point of use.

Mark articles with the conditions that make them unsuitable, such as a plan tier or a region. Reviewers can then check whether the agent respected the boundary. A knowledge base that admits its limits is more trustworthy than one that implies every page fits every case. Review the conditions when a plan or region changes, because an old boundary can block a valid use.

Working with the client on content

Ask the client which topics are sensitive and who must approve an article before agents use it. A policy answer that the client has not approved can create a commitment the desk cannot keep. Record the approval with the article and the date. When the policy changes, the approval expires with it. Record the approval in the article metadata so a reviewer can see who signed off and when.

Share the failed-search list with the client's product and content owners. Some gaps belong to the desk, while others belong to a help centre the client controls. Separating the two prevents the desk from writing around a problem the client should fix. The list is a shared agenda, not a complaint. Agree which owner will respond to each item and when the list will be reviewed again.

Limits and a practical standard

Search data cannot show whether an agent understood an article or whether the customer followed it. It also misses the questions agents answer from memory without searching. Use the data to find candidates for review, then read the work. A metric that is never read alongside a ticket will drift from the service. It also cannot show whether the customer used the answer successfully, which is the outcome that matters.

A practical standard is a defined success signal per request type, a weekly review of failed terms, a sample of article use against real tickets, and a named content owner. Review the measures when the search platform changes. The aim is a knowledge base that agents trust enough to open under pressure. Add one owner for the failed-term review, and keep the last four weekly lists so a recurring gap is visible instead of rediscovered. Keep the failed-term list and the content changes together so the work can be reviewed as a whole, and date each list so a later reader knows which searches produced it.

Sources

  1. HDI support centre industry resources - Industry guidance on support operations, knowledge, and service management practices.
  2. NIST/SEMATECH e-Handbook of Statistical Methods - Measurement and sampling concepts for interpreting usage data.
  3. ISO 18295-1:2017 customer contact centres - Contact-centre process and service consistency concepts.
  4. BLS Occupational Outlook Handbook: customer service representatives - Occupational description of customer service work and information tools.

Related research

Help desk staffing intake

Define the operating lane before staffing begins.

Share the queue, tools, schedule, approvals, and escalation limits. We will use them to prepare a focused help desk staffing brief.

Contact Us